Research Of Efficient Tracking Based On Sparse Representation

Object tracking is a hot subarea of computer vision, aiming at locating an arbitrarily object in a video by estimating its center location, widths, lengths, etc. It is wildly applied in multiple situations, like security surveillance, modern military, traffic navigation and medical image sequence processing. Now, many research agencies and colleges are focusing on this area, and there have been quite a lot of progress. However, due to multiple challenging factors, how to design a robust object tracking algorithm is still worth researching.In this paper, we first introduce the background, current researching situation and challenging factors of object tracking. Then we review several representative tracking methods, among which methods based on sparse representation are the state-of-the-art ones. Based on this, we focus on designing efficient and effective tracking algorithms with sparse representation. Two algorithms are proposed in this work.The first one is object tracking method based on L2regularized least square (L2-RLS) model. Comparing to the traditional L1-RLS model, the proposed L2-RLS model is much more efficient. We also propose a novel representation model with PCA basis and square templates, which ensures the effectiveness of tracking. In addition, we design a likelihood function and model update scheme by explicitly considering partial occlusion.The second one is tracking based on inverse sparse representation and locally weighted distance metric. We reverse the roles of templates and candidates in the traditional sparsity based tracking method so that we reduce the number of L1minimization problems that need to be solved at each frame from several hundred to only one. In addition, we explore why the Euclidean distance metric lose its accuracy when impulse noise occurs from a local perspective. On that basis, we propose the locally weighted distance metric.For each proposed method, we verify its design trough comparing it with several reference algorithms. The experimental results show that each component of a proposed method makes sense. Also, we run the proposed methods on multiple challenging videos, comparing them with more than ten state-of-the-art tracing algorithms. The experimental results show that our algorithm performs favorably against them in terms of both accuracy and speed.